Statistical Methods in Genetic Association Studies and a Genetic Risk Score for Predictive Modeling of Disease Risk: from Gene Discovery to Translation

Fueled by rapid technological and methodological development, research in the area of genetic epidemiology has brought many successes as well as analytical challenges in identification of genetic risk factors for disease risk. Many have predicted that the detection of heritable disease susceptibility variants could eventually lead to stable models for prediction of disease risk. This dissertation is devoted to the statistical methodological consideration and development in genetic association studies, as well as the use of weighted genetic risk score in predictive modeling of disease risk.;We begin the work by reviewing a brief background into genetic association studies, predictive modeling and relevant analytic challenges. In Chapter 2, we discuss the theoretical relationship between classical multiple linear regression (MLR) and two-stage residual-outcome regression analysis (2SR) in terms of confounding adjustment. We demonstrate that the 2SR would introduce bias and loss of power in the presence of confounding, and thus remind researchers to be cautious in applying the 2SR in genetic association studies. In Chapter 3, we propose an adaptive permutation procedure in testing the significance of susceptibility genetic variants. Results show that the adaptive test is statistically valid as well as computationally feasible in genome-wide association studies (GWAS). Recommendations are made for the implementation of adaptive permutation in real studies. In Chapter 4, we develop an explained-variance based genetic risk score (GRS) for predictive modeling of disease risk. Extensive simulation studies suggest that this new weighted GRS is a robust risk score approach that consistently outperforms simple count and odds ratio based GRS approaches. In Chapter 5, we further explore the performance of GRS approaches in the presence of interactions, including statistical interaction, dependence and linkage disequilibrium (LD). Results emphasize the advantage of weighted GRS even in more complicated settings, motivating its application in predictive modeling. This work concludes with several practical guidelines in genetic association studies and predictive modeling, and discussions of future perspectives.